Research map / Video / Video generation

VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling

NeurIPS 2026 · Video › Video generation · approach: Diffusion-based video editing

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VINCIE-NExT, a video editing framework that routes edits through an image-level in-context demonstration (V→I→I→V chain) with a new position encoding (TDF3D-RoPE) and a test-time Chain-of-Editing inference strategy, trained on image, video, and video-derived chain data.

Problem

Video editing requires costly (source video, instruction, edited video) triplets, while image editing data is abundant; the paper seeks to transfer image editing ability to video without large paired video editing data.

Approach

Decompose video editing into a Video→Image→Image→Video chain, prepend an edited image pair as an in-context visual demonstration within a single diffusion transformer trained with a unified flow-matching loss, and use a dual-frequency 3D RoPE so spatially corresponding patches across shots share positions for pixel-faithful edit propagation; Chain-of-Editing runs the chain as two inference stages.

Key results

• VINCIE-NExT achieves an overall OpenVE-Bench score of 3.08, above all open-source baselines (best OpenVE-Edit 2.49) and close to Runway Aleph (3.65).
• Chain-of-Editing with full data raises the ablation model from 2.53 to 2.65 overall, with Creative Edit gaining about 74% over the pure V2V model.
• Training on V<->I<->I data alone with CoE reaches 1.80 overall, up from 1.34 without CoE; an I2I-only model degrades under CoE (1.14 to 1.05).

Surprising findings

• Applying Chain-of-Editing to an I2I-only model degrades performance because the two-stage format is out of distribution.
• A pure V2V model without decomposition (2.32) scores below OpenVE-Edit (2.49), suggesting the decomposed chain format matters more than data alone.

Limitations

• Weak on Camera Edit (2.06 vs Runway Aleph 4.53) since novel-viewpoint synthesis is outside appearance editing scope.
• Local Add and Background Change remain weaker because they require synthesizing content absent from the source.
• Evaluation relies mainly on an MLLM judge (Gemini 2.5 Pro), and the V2V split excludes the unreleased Creative Edit and Camera Multi-Shot Edit categories, limiting comparability with OpenVE-Edit.

Experimental setup

Datasets: OmniEdit OpenVE OpenVE-Bench

Benchmarks: OpenVE-Bench

Metrics: overall score (1-5) instruction compliance consistency and detail fidelity visual quality and stability human pairwise preference

Baselines: Runway Aleph VACE OmniVideo InsViE Lucy-Edit ICVE DITTO OpenVE-Edit AnyV2V

Builds on: VINCIE OpenVE SD3 Flan-T5 RoPE

Closest papers by meaning

• Transforming Image Editors into Video Editors (NeurIPS 2026 · similarity 0.90)
• VicEdit: Learning to Edit Videos from Visual In-Context Examples (NeurIPS 2026 · similarity 0.89)
• Pathways on the Image Manifold: Image Editing via Video Generation (CVPR 2025 · similarity 0.89)
• CCEdit: Creative and Controllable Video Editing via Diffusion Models (CVPR 2024 · similarity 0.88)
• Learning Action and Reasoning-Centric Image Editing from Videos and Simulation (NeurIPS 2024 · similarity 0.88)
• InsightEdit: Towards Better Instruction Following for Image Editing (CVPR 2025 · similarity 0.88)